acceptodds
Under review as a conference paper at ICLR 2027

Continuous First, Discrete Later: VQ-VAEs Without Dimensional Collapse

Abstract

While many approaches to improve VQ-VAE performance focus on codebook size and utilization, the effect of dimensional collapse, where trained VQ-VAE representations live in an extremely low-dimensional subspace (1-2% of full rank), remains poorly understood. We show theoretically and empirically that dimension collapse causes a hard loss lower bound that various codebook improvement techniques fail to surpass. Our analytic framework extends the sequential learning effect of Saxe et al. [2014] by introducing ideas from rate-distortion theory and explains how the latent collapse is caused by the VQ suppressing lower-variance directions. Our theory justifies a simple solution: a "warm-up phase" that trains the model as an (unquantized) autoencoder before introducing VQ. On both synthetic experiments and large-scale image (VQGAN) and audio (WavTokenizer) VQ-VAEs, we show that AE Warm-Up successfully restores representation dimension, leading to lower reconstruction and perceptual loss at the same training budget. Across codebook sizes , AE warm-up raises VQGAN codebook effective dimension from 3-5 to 17-19 and reduces rFID by 17-35%; on WavTokenizer at , it raises codebook dimension from 4 to 17-19 and improves PESQ by 11-14%. We empirically characterize how warm-up duration governs the achievable final loss. In agreement with experiment, our theoretical analysis predicts the latent dimension as a function of warm-up length, enabling an adaptive criterion for switching from AE Warm-up to VQ-VAE training.

Primary area:
learning theory
Keywords:
vqvae, vq-vae, dimension collapse, rate-distortion, water-filling, representation learning, audio tokenization, image tokenization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rEs2OpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates dimensional collapse in VQ-VAEs, where the learned representations occupy only a small fraction of the available latent dimensions, even when codebook utilization is high. The authors propose RD-AE, a simplified theoretical model that replaces vector quantization with a rate-distortion-optimal stochastic channel. Based on this model, they argue that early quantization suppresses low-variance directions and prevents them from being learned. To address this issue, the authors propose AE Warm-up, which first trains the encoder and decoder without quantization before switching to standard VQ-VAE training. Experiments on VQGAN and WavTokenizer show that this simple strategy increases the effective codebook dimension and improves several reconstruction metrics. The authors also derive theoretical bounds connecting warm-up duration with the final effective dimension and reconstruction loss, and provide empirical results supporting the predicted trends.

Critical strengths

The paper studies an interesting and meaningful problem in VQ-VAE training. As the authors point out, high codebook utilization does not necessarily mean that the learned representations make effective use of the available latent dimensions. I find the attempt to explain this phenomenon through training dynamics and rate-distortion theory interesting. Although the theoretical model is simplified, it provides an intuitive explanation for why introducing quantization too early may prevent some latent directions from being learned. Another strength is the simplicity of the proposed solution. AE Warm-up does not require architectural modifications or additional loss terms, and the experiments show improvements on both image and audio models. These results suggest that the proposed training strategy can be useful for conventional VQ-VAE architectures.

Critical weaknesses

My first concern is whether the theoretical analysis sufficiently explains dimensional collapse in real VQ-VAEs, and whether increasing the effective dimension is necessarily beneficial. RD-AE relies on several strong assumptions, including linear encoder-decoder networks, Gaussian inputs, and a rate-distortion-optimal quantization channel. However, real VQ-VAEs involve finite codebooks, imperfect codebook updates, STE approximation errors, and other optimization issues that may also contribute to dimensional collapse. The paper shows that the simplified model exhibits behavior similar to real VQ-VAEs, but this does not establish that the proposed mechanism is the main cause of collapse in practical models. More importantly, I am not convinced that a higher effective codebook dimension should always be considered desirable. For example, Control-GIC (Li et al., 2025) deliberately uses a 4-dimensional codebook while achieving competitive image compression performance. Although this is not directly equivalent to dimensional collapse in a higher-dimensional codebook, it suggests that low-dimensional representations can still be effective. Even in this paper's Table 2, AE Warm-up increases the effective dimension from 4 to 11 at K=4096K=4096, but PESQ, UTMOS, and STOI all become worse. This suggests that increasing the effective dimension does not necessarily improve reconstruction quality. The paper would be more convincing if it could clarify under what conditions dimensional collapse is actually harmful, rather than treating low effective dimension as a general problem.

My second concern is that the experimental evaluation is not sufficiently strong to support the broader practical claims. The image experiments mainly rely on the original VQGAN architecture, while considerably stronger image tokenizers have been developed in recent years. Even after AE Warm-up, the reported ImageNet-1K rFID is 5.35, compared with the publicly reported 4.98 for the original VQGAN (Esser et al., 2021). These numbers may not be directly comparable due to differences in training or evaluation, but the paper does not clarify this gap. It is therefore unclear whether the proposed method remains effective on stronger and more modern backbones. There are also concerns about experimental fairness. For example, Appendix B.2 describes different codebook initialization and dead-code respawn settings in the WavTokenizer training protocols, making it difficult to isolate the contribution of AE Warm-up itself. Finally, I find the connection to actual rate-distortion performance insufficiently explored. Although rate-distortion theory is central to the theoretical analysis, the practical experiments mainly evaluate reconstruction quality at fixed codebook sizes, without reporting actual bitrate or token entropy. Increasing codebook utilization and effective dimension may also change the entropy of the discrete representations, so better reconstruction quality at the same codebook size does not necessarily imply better compression efficiency at the same bitrate. Recent works such as HyperVQ (Yi et al., 2026) and RDVQ (Jiang et al., 2026) have explicitly explored entropy modeling and rate-distortion optimization for VQ-based compression. I am not suggesting that the authors must directly compare against these methods, since their research goals are somewhat different. However, with growing interest in entropy-aware and rate-distortion-optimized VQ methods, I believe it is important to discuss how AE Warm-up relates to these developments and whether improvements in effective dimensionality translate into better compression efficiency. This would help clarify the practical significance of the proposed approach.

References

[1] Li, A., et al. (2025). Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation. International Conference on Learning Representations (ICLR).

[2] Esser, P., Rombach, R., & Ommer, B. (2021). Taming Transformers for High-Resolution Image Synthesis. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

[3] Yi, N., et al. (2026). HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression. Advances in Neural Information Processing Systems (NeurIPS).

[4] Jiang, S., et al. (2026). Differentiable Vector Quantization for Rate-Distortion Optimization of Generative Image Compression. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

Questions for authors
  1. Theoretical explanation of dimensional collapse: The RD-AE analysis is based on a rate-distortion-optimal stochastic channel and simplified linear-Gaussian assumptions. However, actual VQ-VAE training involves finite codebooks, codebook update dynamics, and STE approximation errors. Is there any direct evidence showing that the dimensional collapse observed in real VQ-VAEs is mainly caused by the mechanism identified in RD-AE? For example, have the authors examined whether low-variance directions in real models actually stop receiving useful gradients after quantization, as suggested by the theoretical analysis?

  2. Is a higher effective dimension always beneficial? The paper generally treats increasing the effective codebook dimension as an improvement, but some existing methods, such as Control-GIC, deliberately use low-dimensional codebooks while achieving strong reconstruction performance. Could the authors clarify under what conditions a higher effective dimension is actually necessary? Also, how do the authors explain the results at K=4096K=4096 in Table 2, where AE Warm-up increases the effective dimension but makes several perceptual and speech quality metrics worse?

  3. Experimental strength and fairness: Does AE Warm-up still provide meaningful improvements on stronger and more modern image tokenizer backbones? The best reported VQGAN rFID is 5.35, while the publicly released original VQGAN checkpoint reports 4.98. Could the authors clarify the differences in training and evaluation settings and explain this gap? In addition, could the authors clarify whether the codebook initialization and respawn settings in the WavTokenizer experiments are fully controlled, and how much of the improvement can be attributed to AE Warm-up alone?

  4. Actual rate-distortion performance: Since the theoretical analysis relies heavily on rate-distortion theory, have the authors examined how AE Warm-up affects the entropy or actual bitrate of the learned discrete tokens? Models with the same codebook size can have different token distributions and coding costs. Could the authors provide token entropy statistics from their existing experiments, or clarify whether the improvements would still hold under an explicit rate constraint? Relatedly, how do the authors view the connection between AE Warm-up and recent entropy-aware VQ methods such as HyperVQ and RDVQ?

3 comments

Thank you for the detailed review. We answer your four questions, which should cover the weaknesses you raise.

Q1. Does the RD-AE mechanism explain collapse in real VQ-VAEs?

We treat RD-AE as an idealized model under which the collapse still occurs. In principle, any modeling choice is limited, and we believe that the strength of our setup is to be able to model the observed practical effect rigorously and then translate that into useful predictions (warm up method and water-filling for dimension prediction). The effects you list are not needed to reproduce collapse. Note that we do not dismiss components of the complete system:

  • Finite codebook. It enters as the rate constraint I(z;zq)≤log⁡2KI(z;z_q)\le\log_2 K, which every KK-codeword quantizer satisfies.
  • Imperfect codebook updates. The RD-optimal channel is the best quantizer at given rate, modeling strong codebook.
  • STE. The straight-through estimator and the stop-gradient commitment term are modeled exactly (eqs. 11–12). Their interaction is what freezes the modes below the water level (Sec. 3.2).
  • Linear networks, Gaussian inputs. This is where the training dynamics, including sequential learning, can be solved analytically (Saxe et al., 2014). Both modeled effects are visible in VQGAN and WavTokenizer: sequential learning in their autoencoders, and collapse once VQ is on (Fig. 7).

As a result of our model, we obtain the intuitive explanation that collapse occurs because early quantization combined with sequential learning kills off weak representations (through water-filling picture). To support usefullness of our theory, the predictions we obtain carry over to both practical models. Without warm-up, the codebook dimension stays at 2–5 for every KK (Tables 1–3). Warm-up raises it monotonically as TwuT_{\rm wu} grows (Table 1, Fig. 5). Water-filling on the AE latent spectrum at the switch is tracking the final codebook dimension at every warm-up checkpoint (App. C.1, Fig. 11).

Moreover, the theory leads us o design the fix (warm-up) in the first place, that works in practice. With the quantizer, KK and β\beta unchanged, warm-up raises the VQGAN codebook dimension from 3–5 to 17–19 and lowers rFID by 17–35% (Table 1).

Finally, it builds a general perspective for the field. There are other practical works that play around with rate constraint in the system. For example, the Mimi codec quantizes only half of its training sequences and passes unquantized embeddings to the decoder otherwise (Défossez et al., 2024). Our analysis suggests why this helps: a weaker rate constraint early lets more modes be learned before quantization freezes them.

Q2. When is a higher effective dimension beneficial?

When the codebook is large. Low dimension hurts with growing codebook size, because adding codes then stops lowering the loss. Without warm-up, VQGAN reconstruction improves only modestly as KK grows 64-fold (App. C, Fig. 9). With warm-up the gain grows with KK: rFID improves by 17%, 27% and 35% at K=210K=2^{10}, 2142^{14} and 2162^{16} (Table 1).

Control-GIC is a good example of the setting where choosing higher dimension is not required (though it might be beneficial). It uses K=1024K=1024 codes of dimension 4 (their Sec. 4.1), which is on the lower end of codebook sizes we study. We will cite it as an example of a setting where a low-dimensional codebook is appropriate.

Table 2 shows the same trend for audio, including at K=4096K=4096. This is again a smaller codebook size case. At K=4096K=4096 both arms use the full codebook and are close. Warm-up has the lower Mel loss and is 0.5–2.6% lower on PESQ, UTMOS and STOI. But as KK grows, cold start stops improving and warm-up keeps improving, up to PESQ 2.63 at K=65536K=65536, 12% above the best cold-start run.

So to answer this question, higher dimension is useful at larger codebook size. This is supported by our experiments on VQGAN and WavTokenizer. Most existing research does not address the geometry of latents directly, primarily focusing on growing the codebook size at full utilization, which we show to be suboptimal without resolving the dimensional collapse.

LlamaGen shows that this training "failure" can influence practical decisions. Its ablation settles on codebook dimension 8 and size 16384 because the alternatives fail (their Table 2):

  • dimension 4 is much worse (rFID 9.88 against 2.19);
  • but larger dimensions lose codebook usage (20.9% at dimension 32, 0.29% at 256);
  • doubling KK brings no gain.

Thus they pick dimension 8 at K=16384K=16384 as the best their training reaches. This is the setting our work targets: with warm-up, both KK and the codebook dimension can be raised and keep paying off (Tables 1, 2, 6).

Q3. Experimental strength and fairness

VQGAN baseline In Section 4.1 “VQGAN on ImageNet”, in the “Setup” paragraph, we state that we are comparing training on ImageNet-1K at 500k training steps (~6 epochs), with adversarial loss enabled at 250k steps. In contrast, the released vqgan_imagenet_f16_16384 checkpoint is tagged as step 1,032,743 and epoch 12, and their longer training time primarily explains the difference. We believe that 500k steps is a long enough window to clearly observe that the favorable dynamics of AE Warmup training cause significant gains in loss and rFID over the Vanilla VQGAN and VQGAN w/ Respawn recipe.

Longer training does not remove collapse: the released VQGAN checkpoints use about 4 of their 256 dimensions (Zhang et al., 2025).

Modern tokenizers. For images we use the VQGAN backbone, with several quantizers on top. Warm-up also improves DCVQ (Table 7) and VQGAN-LC at codebook dimension 16 (Table 6). In addition our VQGAN + warmup beats FSQ (Table 5), where warm up does not apply, since the codebook dimension is fixed by design. For audio, WavTokenizer is a recent tokenizer (ICLR 2025). Warm-up acts before quantization and changes only the schedule, so it is easy to apply to other backbones. Because it is a modification supported by our theoretical study with predictions that transfer (ablations on warmup length, water-filling dimension), we believe that combined it is enough to showcase its practical merit.

WavTokenizer settings. Our cold-start baseline uses the WavTokenizer recipe as released, including its k-means codebook initialization and dead-code respawn. We additionally train WavTokenizer at codebook sizes larger than the 4k released by the WavTokenizer authors, to demonstrate that AE Warm-up allows scaling past the 4k codebook size. For the warm-up runs, we scale the dead-code respawn threshold with KK, and use random initialization for our largest codebook size for efficiency.

The obtained cold-start baseline behaves as published: the WavTokenizer authors report the same flat curve when they enlarge the codebook with their recipe (PESQ 2.37, 2.39 and 2.36 at KK = 4096, 8192 and 16384; Ji et al., 2025, Table 5). Our recipe resolves their scaling issue.

For the cleanest estimate of warm-up alone we refer to VQGAN experiments. There, warm-up improves rFID by 17–35%.

Q4. Codebook entropy

Thank you for this question. Our focus is the geometry of the codebook, which token entropy does not measure. As noted in the introduction, prior work mostly raises codebook utilization, which bounds the entropy. However, any tokenizer would benefit from optimal codebook entropy, and in our experiments the entropy is optimal, see table below. Warm-up is complementary: it changes what each bit encodes.

A codebook of size KK costs at most log⁡2K\log_2 K bits per token. Our warm-up runs are close to this ceiling: 13.85 of 14 bits at K=214K=2^{14} on ImageNet-1K.

Token entropy and bitrate on ImageNet-1K (K=214K=2^{14}, code-usage entropy over the full 50k validation set):

ModelStepEntropy HH (bits/token)Perplexity (2H2^H)Codes usedBitrate (bpp)rFID
Vanilla500k9.788809870.038211.37
Vanilla (as in Table 3)850k10.231,1991,3160.04007.91
w/ Respawn500k13.9515,78416,3840.05457.07
TwuT_{\rm wu} = 1k500k13.9415,71416,3840.05457.22
TwuT_{\rm wu} = 6k500k13.8614,88016,3840.05415.97
TwuT_{\rm wu} = 10k500k13.8514,77016,3840.05415.88
TwuT_{\rm wu} = 20k500k13.8314,60416,3840.05405.35
TwuT_{\rm wu} = 30k500k13.8414,69616,3840.05415.37

HyperVQ and RDVQ model and optimize the bits spent on the indices, while warm-up lowers the distortion at a given rate. The two are complementary and potentially could be combined.

Thank you for providing the responses, the training-step clarification in Q3, and the token entropy statistics in Q4. Given that these responses were posted only a few hours before the end of the discussion period, there is unfortunately no time left for multi-round clarification. After carefully reviewing your replies across all threads, my main concerns remain largely unresolved, and several points in the rebuttal raise further questions:

  1. Lack of direct mechanistic evidence and contradiction on codebook dynamics (Q1): Rather than providing direct diagnostics in real VQ-VAEs (e.g., verifying whether low-variance directions actually stop receiving effective gradients after quantization), the response reiterates the RD-AE assumptions and macro-level trends already in the paper. More importantly, in Q1 you state that imperfect codebook updates are not needed to explain collapse because the RD-optimal channel models a strong codebook. However, in your response to Reviewer yPFG (W2), you attribute the failure of β=0\beta=0 to recover collapsed dimensions at K=214K=2^{14} and 2162^{16} to finite codebook update speed—noting that only ~22% and ~6% of codewords receive updates per step, causing codeword staleness. This indicates that the assumption of an always-optimal quantizer breaks down precisely in the large-KK regime (K≥214K \ge 2^{14}) where you argue AE Warm-up is most needed (Q2).

  2. Unexplained degradation at smaller KK and narrowed scope (Q2): While I appreciate clarifying that higher effective dimension is mainly beneficial for large codebooks, the response still does not explain why increasing deffd_{\text{eff}} from 4 to 11 at K=4096K=4096 degrades PESQ, UTMOS, and STOI in Table 2 (or why L1 worsens at longer warm-up durations for K=210K=2^{10} in Table 1). The paper’s framing should be revised to reflect this trade-off rather than treating low effective dimension as universally harmful.

  3. Uncontrolled confounds in WavTokenizer and baseline inconsistencies (Q3):

  • WavTokenizer fairness: Scaling the dead-code respawn threshold inversely with KK (1.0 / 0.5 / 0.25 / 0.06) for AE Warm-up while keeping it fixed at 1.0 for cold-start remains an uncontrolled confound. Furthermore, your claim that the cold-start baseline "behaves as published" (citing flat PESQ of 2.37, 2.39, 2.36 across K=4k,8k,16kK=4\text{k}, 8\text{k}, 16\text{k} in Ji et al., 2025) directly contradicts your own Table 2, where cold-start utilization collapses to 16.9% and 8.4% and test-clean PESQ drops from 2.336 to 2.179 and 2.209. Referring us to VQGAN for "the cleanest estimate of warm-up alone" effectively concedes that the WavTokenizer experiments do not isolate the contribution of AE Warm-up.

  • Image baselines: While the step-count difference (500k vs. ~1M steps) explains the rFID gap to the official VQGAN checkpoint (5.35 vs. 4.98), the new table in Q4 reveals that the Vanilla VQGAN baseline reported in Table 3 (7.91 rFID) was actually trained for 850k steps rather than the 500k-step budget indicated in Figure 5 (11.37 rFID at 500k steps). Evaluation on stronger, modern image tokenizer backbones also remains absent.

  1. Rate-distortion interpretation (Q4): The new table confirms that AE Warm-up improves rFID over the Respawn baseline at comparable token entropy (13.83 vs. 13.95 bits/token). However, achieving near-uniform marginal entropy (~13.85 of 14 bits) reflects full fixed-rate utilization rather than R-D optimality under learned entropy models (e.g., HyperVQ, RDVQ), and incurs a ~35–42% higher bitrate than Vanilla VQGAN (0.0540 vs. 0.0382–0.0400 bpp). A clearer discussion of actual rate-distortion trade-offs remains necessary.

Because the rebuttal not only leaves my core concerns unresolved but also reveals cross-thread theoretical contradictions and confirms uncontrolled confounds in the audio baselines, I am considering lowering my score, if the authors are unable to provide new evidence,

Reviewer_yPFGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explains dimensional collapse in VQ-VAEs using a linear-Gaussian model. The paper attributes the collapse to low-variance directions that are not yet learned when quantization starts. As the water level set by the rate constraint rises, these directions fall below the water level and are frozen. This mechanism gives a theoretical justification for AE warm-up, which trains the model as an autoencoder before quantization is introduced. The paper also gives an upper bound on the number of surviving dimensions as a function of the warm-up length. On VQGAN and WavTokenizer, AE warm-up improves the codebook effective dimension and the reconstruction metrics at the same training budget. AE warm-up also restores the trend that performance improves as the codebook size K increases.

Critical strengths
  1. The paper combines the sequential learning of Saxe et al. (2014) with reverse water-filling. The paper shows that the latent spectrum at the time quantization starts gives an upper bound on the number of surviving dimensions (Theorem 1). Prior work has already given a qualitative explanation that quantization suppresses low-variance directions (Zhang et al., 2025). To my knowledge, turning this explanation into a quantitative theory that depends on the rate and the switching time is new. In the simulations, the same collapse also occurs with dense weights, so the collapse is not an artifact of the diagonal setting. Even in the nonlinear VQGAN, d_eff increases with T_wu, and reverse water-filling on the PCA spectrum of the AE latent gives an upper bound on the final codebook d_eff in all settings. This water-filling diagnostic is cheap to compute and is promising as a guide for choosing when to switch.

  2. Even when codebook utilization is almost 100%, d_eff stays at only a few dimensions, and increasing K gives little improvement. The paper explains this plateau as a problem of the latent dimension. Zhao et al. (2024) have also reported that AE pretraining restores the gains from a larger K. The paper links this recovery of the gains to the recovery of the effective dimension. AE warm-up adds no inference cost and no extra loss term, and improves performance at the same training budget.

Critical weaknesses

As the paper acknowledges, the warm-up approach is not new (Łańcucki et al., 2020; Zhao et al., 2024; and the concurrent work Zhao et al., 2026). The novelty of the paper lies mainly in the theory and in the reframing of the problem. For this reason, the following points have a large effect on my evaluation.

  1. Theorem 1 is informative only when the warm-up is long enough. At T_wu = 0 (cold start), m^wu equals the Shannon-optimal number of modes, and the loss bound reduces to D(R). Therefore, for cold-start collapse, which is the main subject of the paper, the theorem gives only a trivial upper bound on the number of surviving modes. The paper shows the mechanism of the collapse, but shows how severe the collapse is only through simulations. In addition, an upper bound on the dimension alone does not determine the optimal switching time under a fixed budget. The claims of an "adaptive criterion for switching" (Abstract) and of the ability to "adaptively detect the optimal warm-up duration" (Contributions) are therefore overstated. In the experiments, the switching time is chosen by a sweep over T_wu (VQGAN) or by visually checking the plateau of the latent d_eff (WavTokenizer).

  2. The evidence that the RD-AE mechanism is at work in real VQ-VAEs is weak. First, a direct prediction of the theory agrees with the experiments only in part. In the theory, it is the commitment loss that keeps the inactive modes frozen, so the frozen modes should eventually become active again with β = 0. With β = 0 and dead-code respawn, the codebook d_eff increases from 3 (β = 0.25) to 22 at K = 2^10, but stays at about 5 at K ≥ 2^14. The authors hypothesize that this discrepancy comes from the finite adaptation speed of the codebook, which causes staleness between the codewords and z, but they do not test this hypothesis. Second, there is another possible explanation for the effect of warm-up. Zhao et al. (2024) attribute the effect of AE pretraining to solving the following problem: when the codebook is initialized from the outputs of an untrained encoder, the codes concentrate in a narrow region. None of the experiments in this paper distinguishes this initialization-based explanation from the frozen-mode explanation.

[Minor comments]

  • Details of the proof: The argument leading to Eq. (49) relies on the fact that h(k)=k δk(R)+∑j>kσj2h(k) = k\,\delta_k(R) + \sum_{j>k} \sigma_j^2 is non-increasing in k up to the Shannon-optimal number of modes. This holds here because m^wu does not exceed that number, but it is worth stating explicitly. The claim "λj\lambda_j remains ordered" should also be supported by a short argument.

  • The main text cites Zhang et al. (2025) only as a report of the phenomenon. However, Zhang et al. (2025) already give a qualitative explanation that quantization suppresses low-variance directions. The main text should state clearly what RD-AE adds to this explanation (quantitative predictions that depend on the rate and the switching time). The training schedule is also similar in structure to the KL annealing analyzed by Ichikawa & Hukushima (2024) for linear VAEs: both start from an autoencoder and introduce the rate term later. I think a comparison with this work would make the position of the paper clearer.

References

  • Y. Ichikawa and K. Hukushima. Learning Dynamics in Linear VAE: Posterior Collapse Threshold, Superfluous Latent Space Pitfalls, and Speedup with KL Annealing. AISTATS 2024.
Questions for authors
  1. For cold start (T_wu = 0), the bound in Theorem 1 becomes trivial. In this case, can you give a lower bound or an approximate characterization of the number of surviving modes k∞k_\infty?

  2. What specific procedure does the "adaptive criterion for switching" in the Abstract refer to? Does the procedure use the plateau of d_eff (Sec. 4) or the reverse water-filling prediction (Sec. 3.3), and how does the procedure decide the switching time under a fixed budget? Also, how close is the switching time chosen by this procedure to the best point found by grid search?

  3. Could you run an experiment that distinguishes the frozen-mode explanation from the codebook-initialization explanation (Zhao et al., 2024)?

4 comments

Thank you for the careful review. We will address weaknesses and questions together.

W1. Theorem 1 at cold start

Our main theoretical contribution is the RD-AE model, in which the collapse mechanism can be derived and has a simple intuition: a rising water level freezes the modes below it (eqs. 12–15). Theorem 1 is the rigorous statement about the switch: only modes above the water level of the latent spectrum at the switch can survive. You are right that at Twu=0T_{\rm wu}=0 the bound is simply the Shannon count, because no mode has been lost yet. Numerically, only 3–4 modes survive against a Shannon count of 19–24 at the shortest warm-ups (Fig. 6a), but a closed form proving collapse from cold start is left for future work (Q1). Once the leading modes have activated, the bound is informative: in Fig. 6 it is within one mode of the simulated count from Twu≈4T_{\rm wu}\approx 4 on, long before the Shannon count is reached at Twu≈100T_{\rm wu}\approx 100. On VQGAN AE checkpoints, the same water-filling count is tracking the resulting trained codebook dimension (Fig. 11; Fig. 5, right).

Q2. Adaptive switching The method needs no sweep; our TwuT_{\rm wu} sweeps are ablations to test it. During AE training, switch around the point where the latent effective dimension and the water-filling count at R=log⁡2KR=\log_2 K on the same spectrum stop growing; that count then bounds the codebook dimension for each KK. The results are robust to switch time, so the recipe is easy to adopt in practice. On ImageNet-100 the AE dimension plateaus between 10k and 20k steps (Fig. 7a). Switching anywhere from 10k to 40k ends within 0.006 of the best L1+LPIPS in the grid at every KK, and switching at 20k within 0.002. For scale, the best warm-up improves on Respawn by 0.06–0.10 (Tables 1, 4). On ImageNet-1K, switching at 10k, 20k or 30k gives L1+LPIPS within 0.004 of each other (Table 3). For WavTokenizer, at about a week per run, we applied the rule once and switched at 43 epochs (Fig. 7b); PESQ rose 11–14% over cold start at K=213K=2^{13} and 2142^{14} (Table 2).

W2. Evidence for the mechanism in real VQ-VAEs

Let us address β=0\beta=0. In RD-AE at β=0\beta=0 the encoder weight grows (u˙j=2σj2vj>0\dot u_j=2\sigma_j^2v_j>0) until the mode crosses the water level. The channel then transmits the mode at once, because RD-AE assumes a quantizer "trained fast enough to remain optimal for the evolving statistics of zz" (Sec. 3.1). That prediction therefore relies on a codebook that keeps up with the latent statistics. Each training step produces 4096 latents (64 images × 64 tokens; App. B.1), about 4 per codeword at K=210K=2^{10} but only 0.25 at 2142^{14} and 0.06 at 2162^{16}. Even with perfectly even usage, 98% of codewords receive a latent in a given step at K=210K=2^{10}, against about 22% at 2142^{14} and 6% at 2162^{16}, and a codeword that receives none gets no update from that step's data. Setting β=0\beta=0 raises the codebook dimension from 3 to 22 at K=210K=2^{10}, but it ends at 5 at 2142^{14} and 2162^{16}, against 3 and 5 at β=0.25\beta=0.25 (Table 4). The dimension thus recovers only where codewords are updated at almost every step, so β=0\beta=0 is consistent with the theory once the codebook update speed is taken into account. We do not focus on that case, because it offers no new insight into the dimensional collapse we care about.

Q3. Codebook initialization (Zhao et al., 2024).

Zhao et al. (2024) use AE warmup to solve a problem that they call “token collapse”: using VQ from initialization results in a codebook that has many near-0 codewords, while adding VQ to a trained encoder fixes the problem. They attribute this to the fact that a newly initialized VQ-VAE has encoder outputs that are concentrated within a small range (between -.10 and +.30). There are several confounding factors in this observation that differentiate this problem from the one that we are trying to address. Primarily, they train without dead code respawn, so that codes that are initialized to the near-0 region never get updated to keep up with the encoder’s output, resulting in the signature “token collapse” problem they present in Figure 3, where there is a high concentration of codewords in the near-0 region, where the initial untrained encoder distribution was. In our paper, we compare AE Warm-up against baselines with and without dead codeword respawn, denoted in Figure 5 and Table 1 as “VQGAN w/ Respawn” and “Vanilla VQGAN” respectively. Adding dead-code respawn fixes token collapse, as shown in a new figure that we have added to Appendix C.4 and Figure 13. However, as our Figure 5 and Table 1 show, fixing token collapse does not immediately fix dimension collapse, and the model still benefits from AE Warm-up.

Other than that, presented evidence in favour of our model are:

  1. Assumptions. We pick a set of assumptions to model VQVAE, that are designed for analytic tractability of the setup, but they are not detached from real VQVAEs -- every component of the architecture has a corresponding component in our model. This includes rate constraint (through stochastic channel), STE, commitment loss and stop-gradient (that we write down directly), and linear Gaussian setting (to analytically model famous sequential learning phenomena, observed numerically in our AE runs). Combined they create an idealized model (RD-AE) that is both tractable mathematically and reproduces the dimensional collapse. The last part is crucial, because through simplifications we arrived to the solvable setting that has the same phenomena.
  2. Predictions Our model allows us to formulate predictions that we test in practical setups -- resulting dimension increase with warmup duration and waterfilling on latents at the switch closely tracking the effective dimension at the end of training. This predictive power is strong evidence of the model being useful.
  3. Offered Explanation Our model offers an intuitive explanation of the mechanism that goes beyong the assumptions: that collapse occurs because early quantization combined with sequential learning kills off weak representations (through water-filling picture)

Minor comments

Proof details. Both steps are correctly stated; the short arguments are as follows. (i) With f(k)=k δk(R)+∑j>kσj2f(k)=k\,\delta_k(R)+\sum_{j>k}\sigma_j^2, AM–GM applied to δk+1k+1=δkk σk+12\delta_{k+1}^{k+1}=\delta_k^k\,\sigma_{k+1}^2 gives (k+1) δk+1≤k δk+σk+12(k+1)\,\delta_{k+1}\le k\,\delta_k+\sigma_{k+1}^2. So ff is non-increasing in kk on all of {1,…,d}\{1,\dots,d\}, not only up to the Shannon count, and (49) follows from k∞≤mwuk_\infty\le m^{\rm wu}. (ii) On the balanced manifold, (13) gives λ˙j=4(λj−D⋆)+(σj2−λj)\dot\lambda_j=4(\lambda_j-D^\star)+(\sigma_j^2-\lambda_j). For i<ji<j the difference λi−λj\lambda_i-\lambda_j then satisfies a linear differential inequality, so by Grönwall the strict order at TwuT_{\rm wu} is preserved and the active set stays a prefix.

Zhang et al. (2025) attribute collapse to quantization suppressing low-variance directions, and conclude that quantization compresses the spectrum "even when the encoder produces high-rank features" (their Sec. 3.4), making low dimension intrinsic to VQ-VAEs. This argument cannot explain why, with the same quantizer, KK and β\beta, cold start gives 3–5 dimensions and warm-up gives 17–19 of 256 in VQGAN (Table 1). In our study we combine quantization suppression with sequential learning to show, rigorously within the RD-AE model, that early quantization suppresses too many dimensions. The water-filling mechanism and the evolution of the latent spectrum give a bound that depends on the rate and the switching time (Theorem 1), and a theoretical justification of AE warm-up, which we then test and confirm experimentally.

Ichikawa & Hukushima (2024). In their linear VAE with one latent dimension, tanh KL annealing leaves the steady state unchanged and changes only the convergence time (their Thm. 5.4). In RD-AE the switching time changes the steady state (Fig. 6) and so limits the final performance. The studied models differ, but both results favor introducing the rate term later in training as a general principle.

Questions

Q1. A cold-start lower bound, or a closed form for the surviving count, is left for future work. Based on our progress, it is very technically heavy, but it does not lead to any new understanding of the phenomena beyond what is presented in our work.

Q2. See W1: the procedure uses both signals, switching once the AE latent deffd_{\rm eff} and the water-filling count stop growing, and under a fixed budget it gives all remaining steps to VQ. On ImageNet-100, switching at 20k ends within 0.002 of the best L1+LPIPS in the grid at every KK, and switching anywhere in 10k–40k within 0.006.

Q3. See W2 for full answer, but to add to the comparison: RD-AE has no codebook initialization at all: the channel is rate-distortion optimal at every step, and short warm-ups still keep only 3–4 modes (Fig. 6a). The frozen-mode mechanism thus produces collapse with optimal initialization. Whereas resolving initialization issue they outline is possible without warmup, it is enough to apply proper respawning or any other developed for that purpose technique.

Reviewer_yPFGOpenReview ↗

Thank you for the responses; remaining concerns

Thank you for the detailed responses. Since little time remains in the discussion period, further response is not required, but brief comments are welcome if you disagree with any point.

[Points I agree with]

The additional arguments for the proof are both valid, and the AM–GM argument shows that the condition I stated in my review is not needed. I think the comparison with Ichikawa & Hukushima(2024) is also useful (in their setting the steady state does not change, while in RD-AE the switching time changes the steady state). I think these points, and what RD-AE adds to Zhang et al.(2025) (quantitative predictions), should be included in the main text or the appendix of a revised version.

[Remaining points]

First, at cold start Theorem 1 gives only a trivial upper bound on the number of surviving modes, and the severity of the collapse is shown only by simulations. Please also state this in the main text. The results in the response show that the performance hardly changes over a wide range of switching times, not that the proposed procedure finds the optimal time. Please weaken "adaptive criterion"(Abstract), "adaptively detect the optimal warm-up duration"(Contributions), and "on-the-fly automatic decision"(Sec.3.3), so that they describe the procedure as an empirical guideline.

Second, the explanation for β=0 results (finite codebook adaptation speed) has not been tested, and it also implies that the assumption of an always-optimal quantizer does not hold at large K. Also, in the experiments with real models, warm-up changes both the state of the encoder and the codebook initialization at the same time, so the experiments do not separate the frozen-mode explanation from the initialization-based one. The response argues, with the new App.C.4, that dead-code respawn resolves the initialization problem. However, in the Respawn baseline, the codebook is still initialized by k-means on the outputs of the untrained encoder (App.B.1). Respawn improves code usage, but Zhao et al.(2026) argue that the problem caused by this initialization remains even when code usage is healthy. Please state these two points as limitations.

I keep my rating. RD-AE and the upper bound that depends on the switching time are a substantial contribution, and the empirical improvements are confirmed in several settings. However, the response does not provide new results that resolve the two remaining points, which affect my evaluation. The claims can be made consistent with the evidence by revising the text, without new experiments.

Reviewer_CXCAOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper focuses on the effect of dimensional collapse in VQ-VAE training process. The author found that if VQ-VAE performs discrete quantization too early, it can prevent the model from learning rich continuous representations; training a continuous autoencoder first and then enabling quantization can help prevent dimension collapse and improve reconstruction quality. Based on this observation, the authors introduce RD-AE, a toy model for studying VQ-VAE training dynamics, and derive a bound on the latent dimensionality and reconstruction loss as a function of warm-up duration, allowing them to adaptively detect the optimal warm-up duration.

Critical strengths

This paper provides a theoretical and empirical analysis of dimensional collapse in VQ-VAE training process. Based on this analysis, the authors introduce RD-AE, a toy model for studying VQ-VAE training dynamics, and derive a bound on the latent dimensionality and reconstruction loss as a function of warm-up duration, allowing them to adaptively detect the optimal warm-up duration. This method can effectively improve the performance of VQGAN(ImageNet-1K) and WavTokenizer(LibriTTS).

Critical weaknesses
  1. The theory has a limited scope. Strictly speaking, it's developed under conditions such as linear networks and Gaussian inputs, mainly offering a mechanistic explanation for real deep networks, but it can't accurately predict their final loss.
  2. Please provide the experimental results with different initial random seeds, and provide the results with mean and variance values.
Questions for authors
  1. For theoretical analysis, I know the linear networks and Gaussian inputs may be the key requirements for derivation. I hope the authors can provide some toy experiments to verify the generalization of this method by using CNN/Transformer networks and non-Gaussian inputs.
1 comment

Thank you for your feedback! Let us address raised points and questions.

Seeds. The simulations of Section 3 use 128 random seeds: each curve in Fig. 4 is the median, and the cross-seed variance is below 10-3. For VQGAN we already have a second seed on ImageNet-100, and the warm-up gain is stable across the two. The rFID improvement over Respawn at Twu=40T_{\rm wu}=40k is 12.9/19.1/24.0 in Table 1 and 15.5/18.2/23.2 in Table 5, for K=210/214/216K=2^{10}/2^{14}/2^{16} (App. C.3). We will report the mean and standard deviation over several seeds for the experiments in the revision.

Nonlinear networks. We agree that the linear-Gaussian theory explains the mechanism and does not predict the final loss of a deep network (Sec. 5). Its predictions, however, are tested in the setting you ask about. Section 4 trains convolutional networks with attention on real, non-Gaussian data: VQGAN, with self-attention at 16×16 (App. B.1), on ImageNet, and WavTokenizer, with attention in the decoder, on LibriTTS speech. Each step of the mechanism is visible there:

1. Sequential learning. Trained as plain autoencoders, both models show the latent effective dimension dipping and then growing as directions are learned, as in Saxe et al. (Fig. 7).

2. Collapse. With VQ from the start, the codebook dimension stays at 2–5 for every K (Tables 1–3, Fig. 7). Warm-up. Delaying VQ monotonically raises the dimension with increasing TwuT_{\rm wu}, from 3–5 to 17–19, and lowers rFID by 17–35% (Table 1, Fig. 5).

3. Water-filling. Applied to the PCA spectrum of the AE latent at the switch, the theory's water-filling construction tracks the final codebook dimension at every warm-up checkpoint (Fig. 5 right; App. C.1, Fig. 11).

The effect also carries over to another quantizer on the same backbone: warm-up raises the codebook dimension of DCVQ from 8 to 16 and lowers its rFID from 49.3 to 40.2 (App. C.3, Table 7).

Comments on the paper

1 comment

We thank the reviewers; their specific questions are answered in each thread. In short:

Phenomenon. VQ-VAEs can use every codeword yet span only a few latent dimensions, limiting their performance gains from larger codebooks. Clasically studied codebook utilization/entropy miss this.

Theory. We propose an idealized model, RD-AE, that keeps the VQ-VAE ingredients (rate constraint, straight-through estimator, commitment with stop-gradient), exhibits dimensional collapse, and is still analytically solvable. We use it to explain the collapse mechanism in real networks and to make useful predictions.

Mechanism. This leads to us formulating the following intuition. Networks learn strong representations first. A rate constraint imposed early sets a water level, growing rapidly as strong representation learn, and the directions below it become untrained. As a result, VQGAN uses only about 4 of its 256 latent dimensions. An AE, in contrast, eventually learns the weaker directions too, and quantizing it after that, at the same rate, would keep nearly 20.

Empirical validation. Our rigorous model leads to designing AE warm-up: let the representations form first, then quantize. On VQGAN and WavTokenizer it raises the codebook dimension several-fold, improves reconstruction at the same budget (about a quarter lower rFID on ImageNet-1K), and makes larger codebooks pay off again.

Reviews and discussion. The reviewers find the problem meaningful and warm-up simple and effective on both images and audio; rEs2 calls the explanation intuitive, and yPFG notes that a quantitative theory depending on the rate and the switching time is new. The concerns were about scope: whether a linear-Gaussian model explains real VQ-VAEs, how to choose the switch, when a higher dimension helps, and whether codebook initialization could explain the effect instead. We answered them with results already in the paper: the theory's predictions hold in practical VQGAN and WavTokenizer, we provide the way to choose the switch time(and show in ablations that the outcome is robust to the switch time, making the algorithm easy to use), the gain we get grows with codebook size (making this method useful when scaling up the codebook), and testing against dead-code respawn, which removes the initialization problem, but does not remove collapse.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.